HETEROGENEOUS IMPACTS OF CONDITIONAL …the impacts of conditional cash transfers on employment,...
Transcript of HETEROGENEOUS IMPACTS OF CONDITIONAL …the impacts of conditional cash transfers on employment,...
HETEROGENEOUS IMPACTS OF CONDITIONAL CASH TRANSFERS:
EVIDENCE FROM NICARAGUA*
Ana C. Dammert Department of Economics, McMaster University
1280 Main Street W, Hamilton, ON L8S 4M4, Canada E-mail: [email protected]
Abstract
In the last decade, the most popular policy tool used to increase human capital in developing countries has been the conditional cash transfer program. A large literature has shown significant mean impacts on schooling, health, and child labor. Given that transfers depend on regular school attendance and use of preventive health care services, theory predicts differential effects on household behavior. This paper examines heterogeneous effects using random-assignment data from the Red de Proteccion Social (RPS) in rural Nicaragua. Using interactions between the targeting criteria and the treatment indicator, estimates suggest that children located in more impoverished localities experienced a larger impact of the program on schooling and a small impact on the probability of engaging in labor activities. Estimated quantile treatment effects indicate that the there is considerable heterogeneity in the impacts of RPS on the distribution of total and food expenditures. In particular, households at the lower end of the expenditure distribution experienced a smaller increase in expenditures. Keywords: Nicaragua, conditional cash transfers, quantile treatment effect JEL Classification: O15, I38
Current version: September 2007.
2
I. INTRODUCTION The most popular policy tool used in the last decade to increase human capital has
been the conditional cash transfer program, which provides cash payments to households
upon compliance with a defined set of requirements such as regular attendance to school and
visiting health clinics. Many governments implemented experimental frameworks to assess
the impacts of conditional cash transfers on employment, schooling, and health among poor
eligible households (PROGRESA in Mexico, Bolsa Escola in Brazil, and PRAF in Honduras,
among others). Though conditional cash transfers have achieved quantified success in
reaching the poor and bringing about short-term improvements in consumption, education,
and health (Schultz 2004; Gertler 2004; Rawlings and Rubio 2003), most of the literature has
focused on mean impacts. As Heckman, Smith, and Clements (1997) point out, however,
judgments about the “success” of a social program should depend on more than the average
impact. For example, it may be of interested to investigate whether social programs have
differential effects for any subpopulation defined by covariates, for example gender effects,
or whether there is heterogeneity in the effect of treatment. Knowledge of whether a
program’s impacts are concentrated among a few individuals is important for the
effectiveness of the program in reaching the target population.
This paper contributes to the small but growing literature on the estimation of
heterogeneous effects of social programs in developing countries, in particular one type of
program, the conditional cash transfer. This type of program has received a great deal of
attention among policymakers, influencing adoption of new policies in Latin and Central
America. The assessment of heterogeneous impacts is done with a unique data set from a
social experiment in Nicaragua designed to evaluate a conditional cash transfer program
targeted to poor rural households, the Red de Proteccion Social (RPS) or Social Safety Net.
All households in participant RPS localities that satisfy eligibility rules and comply with its
3
requirements receive treatment. The analysis takes advantage of the random assignment of
localities to treatment and control groups so that program participation is not correlated in
expectation with either observed or unobserved individual characteristics and outcome
differences represent the true mean impact of the program.
Conditional cash programs, such as the RPS, have differential effects on household
behavior given that transfers depend on regular school attendance and health visits. For
example, the school cash transfer is conditional on regular attendance of children age 7 to 13
years who did not complete the 4th grade. Households with children in this age category who
are currently attending school will not incur any extra cost and will receive monetary food
and school transfers. The grant acts as a pure income effect and higher expenditures after the
program are expected. For households with children age 7 to 13 years who have not
completed fourth grade and are not attending school, the program has income effects of the
cash transfer and substitution effects of a lower price of schooling driven by the attendance
requirement. Some households may have to bear the cost of children’s foregone labor
earnings due to the implicit reduction in labor time, in which case the impact on household
expenditures may be negative. The monetary transfer received to buy food (or food cash
transfer), however, has a positive effect on household expenditure. The net effect on
expenditures could be positive or negative. Conditional cash transfers have differential effects
based on whether the household is meeting the requirements prior to the implementation of
the program. Knowing more about this heterogeneity is relevant to anti-poverty policies
(Ravallion 2005).
This paper contributes to the existing state of knowledge in several ways. First, the
existing literature focuses on RPS’s mean impacts, in the full sample and in demographic
subgroups. This paper goes beyond mean impacts and interaction variables and test whether
there are heterogeneous impacts of conditional cash transfers on the distribution of
4
expenditures as the theoretical framework predict. In particular, it analyzes the impacts of the
RPS program on the distribution of total and food expenditures using quantile treatment
effect (QTE) estimation. The QTE is the difference in outcomes at various quantiles of the
treatment and control group distributions, which tell us how the expenditure distribution
changes when we assign RPS treatment randomly. Second, the literature on QTE has been
limited mostly to the US context. Recent papers have used QTE to assess the impacts of
training program on labor outcomes such as Heckman, Smith, and Clements (1997), Black,
Smith, Berger, and Noel (2002), Abadie, Angrist, and Imbens (2002), and Firpo (2007).
Bitler, Gelbach, and Hoynes (2005, 2006) examine the impact of a welfare reform experiment
on earnings and total income. Overall, the main finding is that variation in the impact of
treatment across persons is an important aspect of the evaluation problem. To the best of my
knowledge, Djebbari and Smith’s (2005) study represents the first to analyze heterogeneous
impacts of social programs in a developing country. Third, this paper presents evidence of
rank invariance in the RPS context to help clarify the interpretation of the QTE impacts for
the development literature.
The main results show that impact estimates vary among the eligible population. The
first method interacts the treatment indicator with observable characteristics. Results show
that boys experienced a larger positive impact of the program on schooling and smaller
impact on the probability of engaging in labor activities and hours worked. Results also show
that older children experienced a smaller impact of the program on schooling and
participation in labor activities. There are also differential impacts by whether the child is
living with a male head of household and with education of the head of household. To assess
the effectiveness of the targeting criteria, the analysis considers the interaction between the
treatment indicator and marginality index and household per capita expenditures, separately.
5
Main results show that children located in more impoverished areas experienced a larger
impact on schooling and a smaller impact on working hours.
The second method, not tied to covariates, is the QTE, which allows us to test whether
conditional transfers lead to larger or smaller changes in some parts of the outcome
distribution. Estimates from the quantile regression suggest that the positive program impact
in per capita food expenditures and total per capita expenditures is smaller for households
who are in the lower tail of the expenditure distribution. Estimates show that program
impacts are larger for households who had lower levels of food shares prior to the program.
These findings could not have been revealed using mean impact analysis.
The rest of the paper is organized as follows. Section II presents the RPS Program and
data and Section III outlines the theoretical framework. Section IV outlines the empirical
strategy and this is followed by a discussion of the empirical results in Section V. Section VI
concludes.
II. THE RPS PROGRAM
II.1 Program Structure and Benefits
Nicaragua is a lower-middle income country. With an estimated per capita GDP of
US$817 in 2004, Nicaragua remains the second poorest country in the Latin America and
Caribbean region after Haiti. In 2000, to encourage educational attainment and help
impoverished households in rural households, the Nicaraguan government implemented the
Red de Proteccion Social or Social Safety Net. Phase I of the program started with a budget
of US$11 million, representing approximately 0.2 percent of Nicaragua’s GDP (Maluccio
and Flores 2005). With financial assistance of the Inter-American Development Bank and the
Government of Nicaragua, the RPS program was expanded in 2002 with a US$20 million
budget for an additional coverage of three years. The RPS program was made after the
6
Mexican program PROGRESA (now called Oportunidades): it provided benefits conditional
on school attendance and health checkups and participants were identified using a detailed
targeting process aimed at reached poor population in rural areas.
For Phase I of the RPS, the Government of Nicaragua selected the states of Madriz
and Matagalpa from the northern part of the Central Region. This selection was based on the
states ability to implement the program in terms of institutional and local government
capacity, high poverty levels within the communities, and proximity to the capital of
Nicaragua. In 1998, approximately 80 percent of the rural population in Madriz and
Matagalpa was poor and half was extremely poor.
Targeting of poor households was implemented at the RPS headquarters in two stages.
First, officials selected six municipalities within these two states because of the presence of a
concurrent poverty alleviation program. Second, officials ranked all 59 rural comarcas in the
selected six municipalities using a marginality index constructed from the 1995 National
Population and Housing Census. Comarcas (hereafter called localities) are administrative
areas within municipalities including between one and five small communities averaging 100
households each. This index used locality-level information on the illiteracy rate of persons
over age 5, access to basic infrastructure (running water and sewage), and average family
size. The higher the value of the marginality index, the more impoverished is the area. Based
on this score, localities were selected as eligible and ineligible for the RPS program. Out of
59 localities, 42 eligible rural localities were identified as having a high or very high
marginality index and thus pre-selected for the program.1
Program benefits are conditional income transfers composed of (Table 1):
1. Each eligible household received money to buy food (called food cash transfer) every
other month. In order to receive this transfer, a household member (typically the
mother) is required to attend educational workshops and bring their children under the
7
age of five for preventive health care appointments (include vaccinations and growth
monitoring). Children younger than age two were seen monthly and those between
age two and five every other month. In September 2000, the food transfer was
US$224 a year, representing 13 percent of total annual household expenditures in
beneficiary households before the program.
2. Contingent on enrollment and regular attendance2, each household with children age 7
to 13 who had not completed the fourth grade of primary school received a fixed cash
transfer every other month.3 In addition, for each child in the household enrolled in
school, the household received an annual lump sum transfer for school supplies and
uniforms (called the school supplies transfer).4 In September 2000, the school
attendance transfer and the school supplies transfer were US$112 and US$21,
respectively.
To enforce compliance with program requirements, beneficiaries did not receive the
transfer if they failed to carry out the conditions previously described. Less than 1 percent of
households were expelled during the first two years of delivering transfers, though 5 percent
voluntarily left the program, e.g., by dropping out or migrating out of the program area
(Maluccio and Flores, 2005).
II.2 The Experimental Design and Data
The evaluation design is based on an experiment with randomization of localities into
treatment and control groups. One-half of the 42 localities were randomly selected into the
program. The selection was done at a public event in which the localities were ordered by
their marginality index scores and stratified into seven groups of six localities each. Within
each group, randomization was achieved by blindly drawing one of six colored balls without
8
replacement; the first three were selected in the program and the other three in the control
group.5
All households in selected localities are interviewed before and after the random
assignment. The evaluation dataset consists of panel-data observations for 1,359 households
over 3 rounds of survey (baseline: September 2000, follow-ups: October 2001 and October
2002). Surveys at the individual and household level collected information on socioeconomic
and demographic characteristics such as parental schooling, labor market outcomes, health,
nutrition, house infrastructure, among others.6 Table 2 presents descriptive statistics for each
year. Prior to the program, households spend on average C$3885 or about US$298.9 a year,
where 70 percent is allocated to food consumption. Head of households are on average 44
years old, 86 percent are male, and have on average 1.65 years of schooling. Table 2 also
shows that 77 percent of children aged 7 to 13 years attend school and 15 percent participate
in labor activities for an average of 23.5 hours per week.
The randomization is at the locality level rather than at the household or individual
level. There is a greater chance of observing some non-randomness in terms of differences
between localities selected for the control and treatment groups at the household level prior to
the program. As would be expected from the random assignment process, characteristics of
the treatment group are very similar to the control group for all households and subgroups of
the population (Table 3). T-tests of the equality of means suggest that for the majority of
variables measured prior to the random assignment to the treatment and control groups do not
differ. This suggests that the sample is well balanced across the treatment and control
groups.7
9
III. THEORETICAL FRAMEWORK
This section outlines a model of household decision making in the presence of
conditional cash transfers to get a better understanding of potential heterogeneous impacts of
the RPS program. The model is based on Skoufias and Parker (2001) and Djebbari and Smith
(2005). The analysis begins first by considering household time allocation in the absence of
the conditional cash transfer. Neoclassical models of household decision-making are
commonly employed in this analysis. In this framework, parents make decisions about the
allocation of a child’s schooling time, the time of other household members, and the purchase
of goods and services. Parents will invest in each child’s schooling up to the point where
marginal costs of a child’s time in school equal marginal benefits, considering the
opportunity cost of schooling which are the forgone earnings from work.
The opportunity cost of children’s time is likely to vary with observed characteristics.
For example, it is expected to see gender and age differences in child labor if boys and girls
have different returns to education or older children have a comparative advantage in the
labor market. It is important to note that the RPS does not provide higher payments for
female enrollment in school as in PROGRESA. The main idea was to equalize the incentive
for girls in the face of higher wages, on average, for boys in the labor market. Girls in
secondary school received slightly higher subsidies (by about $2 per month) than boys in
Mexico.
The monetary school and food cash transfers are linked to the school attendance of
children aged 7 to 13, participation in health clinics, and other criteria. If they were not
conditioned, transfers will act as a pure income effect. Conditionality of the transfers results
in changes in the marginal cost of investment in schooling. If children participate in the
program with full compliance of the requirements, time devoted to schooling change:
children now receive transfers for attendance and school supplies but lose wages for the extra
10
time the child devotes to schooling. What matters is the ratio of child’s wage and the
marginal increase in earnings due to the transfer.
Participation and compliance with the RPS program might also affect household behavior
as follows: 8
• Households with no children in the targeted age ranges or with children under age 5
(but without children aged 7 to 13 who have not completed the fourth grade) receive
the food cash transfer. These transfers will have a pure income effect and it is
expected that these households will have higher expenditures after the program.
• Households with children aged 7 to 13 years old who have completed fourth grade at
primary school and are attending school will be eligible to receive food transfers but
not school transfers. Food cash transfers will have a pure income effect and it is
expected that these households will have higher expenditures after the program.
• Households with children aged 7 to 13 years, who have not completed fourth grade
but are attending school, are eligible for both the food and school cash transfers.
These transfers will have a pure income effect and it is expected that these households
will have higher expenditures after the program.
• For households with children aged 7 to 13 years who have not completed fourth grade
and are not attending school, the program combines the income effects of the cash
transfer with the substitution effects of a lower price of schooling driven by the
attendance requirement. Some households may have to bear the cost of children’s
foregone labor earnings due to the implicit reduction in labor time. The grant impact
on household expenditures may be negative for these households. The food cash
transfer will have a positive effect on household expenditure. The net effect on
expenditures could be positive or negative.
11
In sum, the predicted effect on expenditures is heterogeneous. At the top of the
expenditures distribution, or the richest households among the eligible ones, households are
meeting or almost meeting program requirements prior to the program and thus the impacts
will be larger. For some part of the bottom of the expenditures distribution are located
households who are not meeting the requirements (children are not going to school the
minimum required time) and for which the cost of participation is the highest (children’s
contribution is significant), for them the program impacts could be positive or negative.
These households are likely to be the ones who rely greatly on child labor. As Basu and Van
(1998) seminal model show a household will send children to work if adult income or family
income from non-child labor sources becomes very low. In between these extremes, the
effect of the program depends on whether the child is attending school the minimum required
time or not.
IV. EMPIRICAL STRATEGY
The most commonly used estimator provides information about the program mean
impacts. In this context, the implicit common effect assumption is present, where the program
is assumed to have the same impact for all individuals in the sample. The existing literature
on heterogeneity of treatment effects predominantly looks at the impact of the program as if it
varies with observed characteristics or subgroups of the population. In the case of conditional
cash transfers, other papers have found evidence of differential impacts on schooling and
child labor for girls vs. boys, and primary school-age children vs. secondary school-age
children, and by socioeconomic status (for example Maluccio and Flores(2005) for RPS;
Skoufias (2005) for PROGRESA, Schady and Araujo(2006) for Ecuador). In addition,
Behrman, Sengupta, and Todd (2005) find that program impacts differ based on children’s
propensities to attend school in PROGRESA. Filmer and Schady (2005) show heterogeneity
12
of program effects by the aggregate measure of parental education and by distance to school
in Cambodia.
For many questions, though, knowledge of distributional parameters is required, for
example the proportion that benefit from treatment, proportion that gain at least a fixed
amount, or the quantiles of treatment effect. Heckman, Smith, and Clements (1997) and
Heckman and Smith (1995) emphasize that many criteria for the evaluation of social
programs require information on the distribution of the treatment effect. To the best of my
knowledge, only Djebbari and Smith (2005) use quantile treatment effects to study
heterogeneous impacts of social programs in a developing country (Mexican program
PROGRESA). This section describes two preferred econometric estimations used for
obtaining impact estimates that vary among the eligible population.
I.V.1 Impacts at the Subgroup Level
The first method generates impacts estimates that vary among the eligible population.
It considers variation in impacts as a function of observable characteristics through the
interaction of the treatment indicator in equation (1) with a variety of individual and locality
characteristics as follows:
0 1 2 3* *i i i i i i iy C T T C Xβ β β β α ω= + + + + + (1)
where iy is some outcome measure, iC is the characteristic of interest, iT is a dummy
variable representing whether the locality was randomly assigned to the treatment or control
group, and *i iT C represent the interactions between the characteristics and the treatment
indicator. Since the specification does not condition on household participation in the
program, but only on whether the household resides in a locality were the program is
available, the estimates reflect the “intent-to-treat” effect of the program or the effect for
individuals who are offered the treatment.9
13
The interpretation of the coefficients is as follows: for example in the specification
that tests for heterogeneous impacts by gender, iC is a dummy variable representing if the
child is male, the coefficient 1β is an estimate of the difference in the outcome between boys
and girls, the RPS effect for girls is given by 2β , the corresponding effect for boys is given by
the sum of the coefficients 2 3β β+ . If 3β is statistically significant there is evidence of
heterogeneity of treatment effects by gender. iC include also characteristics of the head of
household and locality. I also use a criterion used by PRS to select beneficiaries, the locality
marginality index. Program officials using data from the 1995 Nicaraguan Household Survey,
collected prior to the program, construct this index. Following the analysis in Djebbari and
Smith (2005) for Mexico, the impact of conditional cash transfers is expected to be largest for
households living in more impoverished localities as defined by the marginality index. If the
targeting mechanism is efficient, then households in the most marginal localities get a greater
program impact than less marginal places. Equation (1) also controls for other baseline
household and individual characteristics ( iX ) to take into account any differences that were
present despite randomization and to increase the precision of the coefficient estimates. The
standard errors are clustered at the locality level.
I.V.2 Quantile Treatment Effects
The second econometric method used to derive estimates that vary among the eligible
population is the quantile treatment effect. The second method uses a quantile regression to
provide a convenient framework to characterize the heterogeneous impact of treatment on
different points of the outcome distribution. Most of the existing literature on QTE is based
on social experiments from U.S. employment, training and welfare programs. Heckman,
Smith, and Clemens (1997), find strong evidence that heterogeneity is an important feature of
14
impact distributions using experimental data from the National Job Training Partnership Act
Study (JTPA). Black, Smith, Berger, and Noel (2002) using experimental data from the
Worker Profiling and Reemployment Services program find that the estimated impact of
treatment varies widely across quantiles of the outcome distributions. The pattern of impacts
suggests that the treatment has its largest effect on persons whose probability of
unemployment insurance benefit exhaustion without treatment would be of moderate
duration. In evaluating the economic effects of welfare reform, Bitler, Gelbach, and Hoynes
(2005, 2006) find strong evidence against the common effect assumption using experimental
data from the Connecticut’s Job First Waiver program and the Canadian Self-Sufficiency
Project. Their estimates suggest substantial heterogeneity in the impact of welfare reform on
earnings and total income, which is consistent with the predictions from the static labor
supply model.
Let 1Y and 0Y denote the outcome of interest in the treated and control states with
corresponding cumulative distribution functions 1 1( ) Pr[ ]F y Y y= ≤ and 0 0( ) Pr[ ]F y Y y= ≤ .
Denote the thθ quantiles of each distribution
( ) inf{ : ( ) }, T 0,1Ty T y F yθ θ= ≥ = . (2)
Thus, we can define the quantile treatment effect as ( 1) ( 0)QTE y T y Tθ θ θΔ = = − = . For
example, suppose that y represents family income in a given year, 0.25y is that level of income
for households in the treatment (control) group such that 25 percent of treatment (control)
households have income below it. 0.25
QTEΔ is given by the difference between the income of
households in the 25 percent sample quantile of the treated distribution and the 25 percent
sample quantile of the outcome distribution.
The impact estimate for a given quantile θ distribution is the coefficient on the
treatment indicator from the corresponding quantile regression as follows
15
( | ) ( ) ( ) , (0,1)iiQ y T Tθ α θ β θ θ= + ∈ (3)
where ( | )iQ y Tθ denotes the quantile θ of expenditures conditional on treatment. Following
Koenker and Basset (1978), the thθ quantile estimator can be defined as the solution to the
problem:
' ' 1: :
1 1min | | (1 ) | | min ( )i
i i
n
i i i iii y T i y T
y T y T un nθ θ θ θβ β
β β
θ β θ β ρ=≥
⎡ ⎤− + − − =⎢ ⎥
⎢ ⎥⎣ ⎦∑ ∑ ∑
≺
(4)
where (.)θρ is known as the check function.
As presented above in table 2, the RPS sample is well balanced and there are few
statistical differences in the observable characteristics in the two groups. To correct for any
differences not accounted for the randomization of localities into treatment and control
groups and to obtain more precise estimates I have included covariates as in Djebbari and
Smith (2005) by estimating the QTE on Y , as follows:
1. Estimate the effects of observable characteristics (Xi) on the outcome of interest (Yi) as
1
K
i ii
Y Xα β υ=
= + +∑ .
2. Predict the residuals by removing the effect of household characteristics from the
outcome as 1
[ ]K
i ii
Y y Xα β=
= − +∑ .
3. Estimate the QTE on the residuals Y against the treatment indicator as
( | ) ( ) ( ) , (0,1)iiQ Y T Tθ α θ β θ θ= + ∈ .
Estimating QTE on Y is similar to estimating mean impacts conditioned on X. The
control variables include characteristics of the head of household (age, education, gender,
employment) and household demographic composition.10 The advantage of QTE approach
relative to the common effect model is that the impact of the program on different quantiles
of the outcome distribution does not have to be constant.
16
Note that although average differences equal differences in averages, the treatment
effect at quantile θ is not the quantile of the difference (Y1- Y0). The QTE corresponds, for
any fixed percentile, to the horizontal distance between two cumulative distribution
functions. Under the rank preservation assumption, QTE can be interpreted as the treatment
effect for individuals at particular quantiles of the control group outcome distribution or the
treatment effect for the person located at quantile θ in the distribution (Bitler et al. 2005).
Without the rank preservation assumption, QTE represent how various quantiles of the
outcome distribution change in the treatment and control groups, but we cannot make
inference on the impact on any particular person.
Rank preservation across treatment status is a strong assumption though as it requires
that the rank of the potential outcome for a given individual would be the same under
treatment as under non-treatment. There are two ways to deal with cases where the rank
invariance assumption is not valid, Heckman, Smith, and Clements (1997) suggest computing
bounds for the QTE, allowing for several possibilities of reordering of the ranks. The second
approach argues that even without this assumption, QTE estimates are informative about the
overall impacts of the program and one can still have meaningful parameter for policy
purposes. In the absence of rank invariance, the interpretation of QTE is the difference in the
treated and control distributions, not the treatment effects for identifiable people in either
distribution (Bitler, et al. 2005 and 2006). The last section of the paper analyzes whether the
rank invariance assumption is valid in the RPS context.
Outcomes of Interest
The outcomes of interest in the empirical section include household and individual
level variables. To analyze the impact on household welfare the empirical literature uses
household consumption rather than income because data on expenditures are likely to be
17
more accurate and consumption expenditures have a stronger link with current levels of
welfare (Deaton 1997). At the household level, this paper analyzes three outcomes of interest:
per capita total expenditure, per capita food expenditure, and food share of total expenditures.
The analysis of food expenditures is important because one of the keys of the program is
supplementing income to increase expenditures on food so as to improve household nutrition.
The expenditure variables include food, non-food items, and the value of food produced and
consumed at home.11
The RPS program also aims at improving the educational and health outcomes of
children. I focus on children aged 7 to 13 years old at the baseline because they are most
likely to be affected by the conditionality of the cash transfers. Outcomes of interest include
child labor (participation and working hours) and school attendance. Child labor refers to
children who are engaged in market work, which includes wage employment, self-
employment, agriculture, unpaid work in a family business, and helping on the family farm.12
Impacts are estimated using OLS, except in the case of hours worked where the dependent
variable is censored at zero therefore impacts are estimated using a Tobit model.
V. RESULTS
V.1 Impacts along Observable Characteristics
Table 4 reports the treatment indicator interacted by different covariates of interest.
The main results show that the program has different impacts on children with different
observable characteristics. For example, boys aged 7 to 13 years experienced a larger and
statistically significantly different impact on school attendance than girls. Estimates suggest
that the RPS program increased school attendance by 12 percentage points for girls and by 18
percentage points for boys in 2001. In addition, the reduction in the probability of engaging
in market activities is larger for boys. Estimates show that boys experienced a greater impact
of the program on the probability of engaging in market work and hours worked. Results
18
show that the RPS program decreased participation in labor activities for boys by 11
percentage points in 2001 and 14 percentage points in 2002, while the negative effect of the
RPS program on labor participation for girls is small (1 percentage point in both years).
These findings are important given that the program did not provide differential transfers to
boys and girls. For instance, PROGRESA provided slightly more money to girls enrolled in
secondary school and the results show that the program had a greater impact on secondary
age girls (Skoufias and Parker 2001). In addition, it is important to note that this definition of
work does not include other activities usually not remunerated and performed in the same
household, such as taking care of younger siblings, cleaning, and cooking, among other
household chores. A broader definition including detailed household chores may decrease
this gender difference in participation rates.
The coefficient on the interaction term between treatment and age shows that older
children experienced a smaller impact of the program on schooling, as well as on the
probability of engaging in market work and hours worked. This is related with previous
findings that with higher age potential earnings increase, thus transfers might not be high
enough to compensate for forgone earnings. The interaction with household head education
shows that children with more educated head of households experienced a smaller impact of
the program on schooling. Empirical literature has shown that higher educated parents use
more efficiently the information provided in health clinics about nutrition and value more
schooling and less child labor (Strauss and Thomas 1996). Thus, school attendance among
children living in more educated households would be higher and the margins for
improvement lower than among children in households with lower parental education. In
addition, children living with male head of households experienced a smaller impact of the
program on school attendance and larger impact on the probability of engaging in labor
activities and hours worked.
19
Finally, to analyze the targeting mechanism of the program the last rows show the
effect of the treatment interacted with the marginality index and household per capita
expenditures separately. Based on the marginality index, I group households into quintiles
and interact the treatment indicator with the index categories. If the actual targeting of the
program is efficient then the impact in schooling should decrease from the poorest index
quantiles to the richest index quantiles. As estimates from Table 4 show, children living in
more impoverished areas experienced larger impacts of the program on school attendance in
2001. Similar results are obtained with the interaction of the treatment indicator and quintiles
of household per capita expenditures. In 2002, however, children living in more
impoverished areas experienced a smaller impact of the program on schooling.
The estimates also show that children in poorest households experienced smaller
impacts of the program on the probability of engaging in labor activities. This is consistent
with the fact that the RPS is a conditional program. Households living in a more
impoverished area face a trade-off between sending children to school and receiving the
school cash transfer or not sending them to school and receiving their labor income or
benefiting from their help at home or in the family farm/business. Results suggest that
children appear as attending school and working part time rather than attending school only.
These results are consistent with estimates using quintiles of per capita household
expenditures.
V.2 Quantile Treatment Effect Regression
The quantile treatment effect provides information on how the impact at the
household level varies at different points of the expenditure distribution. Figure 1 through 6
plots the quantiles using post-treatment data. The solid line represents the estimate of the RPS
treatment in a given quantile. The associated 95% confidence intervals are obtained from the
20
bootstrap with 1000 replications clustered at the locality level. These bootstrap confidence
intervals are plotted on the graph with dashed lines. For comparison purposes, the mean
treatment effect is plotted as a small dashed line.13
Overall, RPS treatment group expenditures are greater than control group
expenditures, yielding positive impacts at each quantile of the distribution. For per capita
total expenditures and per capita food expenditures, the difference increases from the lowest
percentile to the highest percentile of the distribution. These findings suggest that households
with lower expenditures tend to receive lower positive impacts from the program. As the
theoretical framework suggest the impacts are greater for households with higher
expenditures who are more likely meeting or almost meeting program requirements prior to
the program. For households with lower expenditures who are more likely not meeting the
requirements and cost of participation is the highest, program impacts are still positive but
smaller than for households at the upper end of the distribution. These results are similar to
Djebbari and Smith (2005) findings for the Mexican’s PROGRESA, which provides benefits
conditional on beneficiaries fulfilling human capital enhancing requirements: school
enrollment of children aged 8–16, attendance by an adult at a monthly health seminar, and
compliance by all family members to a schedule of preventive health checkups. They find
that program impacts on wealth and nutrition are greater for households who were at higher
levels of wealth and nutrition prior to the program. Similarly, for the share of food
expenditure, the difference decreased from the lowest percentile to the highest percentile
suggesting that the program impacts are higher for households who had lower levels of food
shares prior to the program.14
Figure 1 shows that in 2001 the program impact on per capita total expenditures varies
from about C$707 (US$ 54) to C$3087 (US$ 237). In 2002, the program impact on per capita
total expenditures varies from about C$264 (US$20) to C$1293 (US$99) for the highest
21
percentile (Figure 2). Many of the impacts are quite large compared to the impact at the mean
of C$1184 and C$820 in 2001 and 2002, respectively. These results suggest that households
at the top of the outcome distribution receive more than five times the impact that households
with lower expenditures do.
I can test whether a constant treatment effect could lead to a range as large as that for
the QTE point estimate as in Bitler et al. (2006). The test is as follows, using the bootstrap.
First, keep only observations in the control group and assign a uniformly distributed random
number to the ith household in the bth bootstrap sample. Second, sort the sample of households
using this random number and assign t=1 to households with number the random number
higher than 0.5 and in the bth sample and t=0 to the remaining households in this bootstrap
sample. Third, add the estimated mean treatment effect to households with t=1 to create a
synthetic null treatment group distribution. Finally, use the synthetic null treatment group and
the remaining control group to construct the QTE for the null hypothesis. From the resulting
individual distributions, we can generate a confidence interval for testing the maximum
minus minimum range, which compares the distribution for the range for the null with the
real-data QTE range. This confidence interval is estimated with 99 bootstrap replications.
This test suggest that a confidence interval for the null constant treatment range is [2896.2,
3699.5] and [2694.6, 3500.5] at a confidence level above 95 percent for 2001 and 2002,
respectively. The QTE range estimated using the data is 2380.3 and 1029.8 for 2001 and
2002. These results show that the mean treatment effect is not sufficient to characterize
RPS’s effects on total per capita expenditures.
Consistent with the RPS program’s goal, additional expenditures as a result of the
transfers were spent predominantly on food. Results for food expenditures suggest a large
degree of treatment impact heterogeneity. In 2001, the program impact on per capita food
expenditures varies from about C$367 (US$28) to C$3780 (US$290) for the highest
22
percentile of the distribution (Figure 3). In 2002, the program impact on per capita food
expenditures varies from about C$174 (US$13) to C$1846 (US$142) for the highest
percentile of the distribution (Figure 4). The impact at the mean is C$1004 and C$733 for
2002 and 2001, which are far below the impact at the top of the distribution. The confidence
interval for a null of constant treatment effects is [1928.2, 2549.7] and [1647.5, 2486.8] at a
confidence level of above 95 percent, while the estimated range over all quantiles in the real
data is 3412.2 and 1671.7 for 2001 and 2002 respectively. The positive impact of the program
for households with the highest per capita food expenditures prior to the program is almost
seven times the impact for households with lower food expenditures, which is not captured
by the mean treatment effect estimate.
To further explore the impacts of RPS on the distribution of expenditures, Figure 5
and 6 show QTEs for the share of food expenditures in the household budget. In 2001, the
program impact on food share ranges from about 7.79 percentage points to -0.18 percentage
points for the highest percentile. In 2002, the program impact on food share varies from about
8.65 percentage points to -1.42 percentage points for the highest percentile. The impact at the
mean is about 4.0 and 3.8 percentage points in 2001 and 2002. The impact is higher for
households who have lower share of food expenditures prior to the program. Maluccio and
Flores (2005) have shown that not only the number of food items purchased increased but
also their nutritional value.
Rank Preservation and Rank Reversal
The main QTE findings show that the impact of the RPS program varied across the
distribution of total and food expenditures. As previously discussed, the impact of the
treatment on the distribution is not the distribution of treatment effects. Only under the rank
preservation assumption, this interpretation is valid. This section examines whether there is
23
evidence of rank preservation. As in Bitler et al. (2005) I can use the treatment and control
distributions of demographic characteristics to see if there is evidence against rank
preservation or rank reversal. For example, if the distribution of observable characteristics in
some range of the expenditures distribution varies significantly between the treatment and
control group, this would be evidence against rank preservation. Note, however, that finding
no significant differences in demographics does not imply rank preservation. In addition,
even if observable characteristics do not change rank reversal may have occurred among
unobservables. As in Bitler et al. (2005), the test of rank reversal is as follows. First,
sampling with replacement from the real data and draw 500 bootstrap samples, each having
the same number of observations as the real data. Second, for the bth replication sample,
randomly order the observations and assign t=1 to the first 706 households in the bth sample
and t=0 to the remaining households in this bootstrap sample. Third, calculate the mean of
the observable characteristic in each synthetic group and take the difference of these sample
means (db). Fourth, sort all sample means db from lowest to highest. Finally, use d25 and d475
respectively and construct a 90 percent confidence interval.
A real-data estimated difference is significantly different from zero if it falls outside
the interval just defined. The p-value is defined as 1 ( 1) 501p k= − − , where k is the index
variable, which is estimate as follows: for each synthetic group create a vector with the
sample means from the bootstrap (500 observations) and add the mean difference using real
data, take the absolute value, and sort these observations. Let k be the index of the real data
observation in the sorted data, for example if the real data estimate in the sorted data lies
between the 300 and 301 bootstrap realization, then k = 301.
Tables 5 and 6 show the difference in the demographic variable between the treatment
and control group within a given quantile and their p-value for statistical significance. I
analyze household head characteristics (gender, education, age, and employment) and
24
household demographic composition (girls 0-5 years, boys 0-5 years, girls 6-15 years, and
boys 6-15 years). Panel A classifies people by their position (quantile) in the per capita total
expenditure distribution and Panel B classifies people by their position in the per capita food
expenditure distribution. Table 5 presents the results from the exercise using the 2001 data
whereas Table 6 presents the results using the 2002 data. Of the 128 differences, 18 are
statistically significant at the 10 percent level or below in 2001 and 2002. The individual test
suggests that some rank reversal may be present based on this demographic characteristics.
The joint test for the significance of the differences within a given quantile range suggests
that the chi-square statistic fails to reject the joint test for all ranges of per capita food
expenditure and per capita total expenditure.15 More work, however, would be needed to go
further in discussing rank reversal in the RPS program.
VI. SUMMARY AND CONCLUSIONS
Recent empirical work has shown that conditional programs are effective in raising
school achievements and improving health outcomes in Mexico, Brazil, and Nicaragua. Much
of the overall literature has focused on mean impacts. This study presents evidence on
heterogeneous impacts of conditional cash transfers using a social experiment from a poverty
alleviation program in Nicaragua. One way to analyze heterogeneous treatment impacts is to
estimate interaction terms between the treatment indicator and the observable characteristic.
For children, the results show that boys aged 7 to 13 years experienced a larger impact of the
program on increasing schooling, but boys experienced a smaller impact of the program in
reducing participation in labor activities. In addition, older children experienced a smaller
impact of the program in reducing participation in labor activities and working hours. Finally,
the results show that children living in more impoverished localities experienced larger
impacts of the program on schooling and smaller impacts on working hours.
25
To illustrate heterogeneous impacts not tied to covariates, I use quantile treatment
effect regression. The results suggest evidence against the common effect assumption. For the
RPS program, the estimated treatment effect appears to differ across quantiles of the outcome
distribution. In particular, for per capita total expenditures and per capita food expenditures,
the impact of the program is lower for households who were at a lower level of expenditures
prior to the program. That is, the RPS program has greater effect on households who would
otherwise have had high per capita total and food expenditures. Quantile treatment effect
estimates show there was considerable heterogeneity in the impacts of the RPS on the
distributions of expenditures, which is missed by looking only at average treatment effects.
As the theoretical framework suggest the impacts are greater for households with higher
expenditures who are more likely meeting or almost meeting program requirements prior to
the program. For households with lower expenditures who are more likely not meeting the
requirements and cost of participation is the highest, program impacts are still positive but
smaller than for households at the upper end of the distribution. These findings could not
have been revealed using mean impact analysis. This has important implications for the
implementation and evaluation of conditional cash transfers that are spreading rapidly in
developing countries.
26
REFERENCES
Abadie, Alberto, Joshua Angrist, and Guido W. Imbens. 2002. Instrumental Variables
Estimation of Quantile Treatment Effects. Econometrica 70(1):91-117.
Behrman, Jere R., and Petra Todd. 1999. Randomness in the Experimental Samples of
PROGRESA (Education, Health, and Nutrition Program). Report submitted to
PROGRESA. International Food Policy Research Institute, Washington, D.C.
Behrman, Jere R., Piyali Sengupta, and Petra Todd. 2005. Progressing through PROGRESA:
An Impact Assessment of a School Subsidy Experiment in Rural Mexico. Economic
Development and Cultural Change 54: 237–275
Bitler, Marianne, Jonah Gelbach, and Hilary Hoynes. 2005. Distributional Impacts of the
Self-Sufficiency Project. NBER Working Paper No. 11626. September.
---------------------------------. 2006. What Mean Impact Miss: Distributional Effects of
Welfare Reform Experiments. American Economic Review 96(4):988-1012.
Black, Dan, Jeffrey Smith, Mark C. Berger, and Brett J. Noel. 2003. Is The Threat of
Reemployment Services More Effective Than The Services Themselves?
Experimental Evidence from the UI System. American Economic Review 93: 1313-
1327.
Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach to
Development Policy. Washington D.C: Johns Hopkins University Press.
Djebbari, Habiba, and Jeffrey Smith. 2005. Heterogeneous Program Impacts in PROGRESA.
Manuscript.
Filmer, Deon, and Norbert Schady. 2006. Getting Girls into School: Evidence from a
Scholarship Program in Cambodia. World Bank Policy Research Working Paper
3910. Washington D.C.
27
Firpo, Sergio. 2007. Efficient Semiparametric Estimation of Quantile Treatment Effects.
Econometrica 75(January): 259-276.
Gertler, Paul. 2004 . Do Conditional Cash Transfers Improve Child Health? Evidence from
PROGRESA's Control Randomized Experiment. The American Economic Review
Papers and Proceedings 94(2):336-341.
Heckman, James, Jeffrey Smith, and Nancy Clements. 1997. Making the Most Out of
Programme Evaluations and Social Experiments: Accounting for Heterogeneity in
Program Impacts. The Review of Economic Studies 64: 487-535.
Heckman, James, Robert LaLonde, and Jeffrey Smith. 1999. The Economics and
Econometrics of Active Labor Market Programs in Orley Ashenfelter and David Card
(eds.), Handbook of Labor Economics, Volume 3A. Amsterdam: North-Holland, 1865-
2097.
Heckman, James and Jeffrey Smith. 1995. Assessing the Case for Social
Experiments. Journal of Economic Perspectives 9: 85-110.
Koenker, Roger, and Gilbert Basset. 1978. Regression quantiles. Econometrica 46(1): 33-50.
Maluccio, John, and Rafael Flores. 2005. Impact Evaluation of a Conditional Cash Transfer
Program: The Nicaraguan Red de Proteccion Social, FCND Discussion Paper, No.
141, IFPRI, Washington D.C.
Maluccio, John, Alexis Murphy, and Fernando Regalia. 2006. Does Supply Matter? Initial
Supply Conditions and the Effectiveness of Conditional Cash Transfers for Schooling
in Nicaragua. Inter-American Development Bank. Manuscript.
Ravallion, Martin. 2005. Evaluating Anti-Poverty Programs in T. Paul Schultz and John
Strauss (eds.), Handbook of Development Economics Volume 4. Forthcoming.
28
Rawlings, Laura, and Gloria Rubio. 2003. Evaluating the Impact of Conditional Cash
Transfer Programs: Lessons from Latin America. World Bank Policy Research
Working Paper No. 3119. Washington D.C.
Schady, Norbert, and Maria Caridad Araujo. 2006. Cash Transfers, Conditions, School
Enrollment, and Child Work: Evidence from a Randomized Experiment in Ecuador.
World Bank Policy Research Working Paper 3930.
Schultz, T. Paul. 2004. School Subsidies for the Poor: Evaluating the Mexican Progresa
Poverty Program. Journal of Development Economics 74(1): 199-250.
Skoufias, Emmanuel, and Susan Parker. 2001. Conditional Cash Transfers and Their Impact
on Child Work and Schooling: Evidence from the PROGRESA Program in Mexico.
Economia 2: 45–96.
Skoufias, Emmanuel. 2005. PROGRESA and Its Impacts on the Welfare of Rural
Households in Mexico. IFPRI Research Report No. 139. International Food Policy
Research Institute, Washington, D.C.
29
Table 1: Nicaraguan RPS Beneficiary Requirements
Household Type With no
targeted children
With children aged 0-5
With children aged 7 -13 who
have not completed 4th
grade (A) (B) (C)
Attend bimonthly health education workshops
Bring children to prescheduled health care appointments Monthly ( 0-2 years) Bimonthly (2-5 years)
Adequate weight gain for children under 5a Enrollment in grades 1 to 4 of all targeted children in the household
Regular attendance (85%) of all targeted children in the household
Promotion at end of school year b Teacher transfer Up-to-date vaccination for all children under 5 years
a This requirement was discontinued in Phase II in 2003 b This condition was not enforced Source: Maluccio and Flores (2005)
30
Table 2: Descriptive Statistics for Children and Their Households
2000
2001 2002
Household Level Per capita Total Consumption 3885.08 3852.74 3880.91 Per capita Food Consumption 2672.33 2669.20 2634.18 Food share 0.704 0.688 0.683 Head of Household a Age 44.27 46.05 47.01 Male 0.858 0.858 0.858 Years of Education 1.652 - - Children 7-13 years at baseline
Gender 0.521 0.521 0.521 Age 9.847 10.938 11.969 School Attendance 0.766 0.885 0.855 Participation in labor activities 0.150 0.109 0.177 Weekly Working Hours 3.51 2.90 5.31 Weekly Working Hours conditional on employment 23.47 26.49 30.07 a Data available for 2000 Note: Expenditures levels are in Nicaraguan Cordobas, the equivalent exchange rate is US$1 = C$13. Sample includes 1359 households with observations in the panel 2000-2001-2002.
31
Table 3: RPS Summary Statistics: 2000 Baseline (Standard error in parenthesis)
Treatment
Control Difference
Household Level Per capita Total Consumption 4020.897 3738.242 282.656 (203.22) (212.23) (290.34) Per capita Food Consumption 2759.874 2577.680 182.193 (129.02) (128.33) (179.81) Food share 0.700 0.708 -0.008 (0.01) (0.01) (0.01) Household Head Age 44.637 43.864 0.774 (0.85) (0.73) (1.11) Male 0.868 0.847 0.021 (0.01) (0.01) (0.02) Years of education 1.698 1.602 0.096 (0.14) (0.09) (0.16) N 706 653 1359 Children 7-13 years Gender 0.532 0.509 0.023 (0.02) (0.02) (0.02) Age 9.824 9.874 -0.050 (0.07) (0.07) (0.10) School Attendance 0.766 0.767 -0.001 (0.01) (0.01) (0.02) Participation in labor activities 0.142 0.158 -0.016 (0.01) (0.01) (0.02) Working Hours 3.264 3.782 -0.517 (0.33) (0.37) (0.49) N 916 829 1745 *Statistically significant at 5% level, **Statistically significant at 10% level (only for differences). Note: Expenditures levels are in Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/.13. Robust standard errors, clustered at the locality level. Sample includes 1359 households with observations in the panel 2000-2001-2002
32
Table 4: RPS Program Impacts along Observables Characteristics for children 7 to 13 years at baseline (Standard errors in parenthesis)
2001 2002 School
Attendance Participation
in labor activities
Hours Worked (Tobit)
School Attendance
Participation in labor activities
Hours Worked (Tobit)
T*male 0.060* -0.099* -8.085 0.063* -0.124* -7.330 (0.03) (0.04) (9.27) (0.03) (0.04) (7.59) T 0.117* -0.012 -12.335 0.110* -0.014 -11.778 (0.03) (0.01) (8.52) (0.02) (0.02) (8.74) T*age -0.003 -0.019* -1.677 0.018 0.000 3.712* (0.01) (0.01) (1.91) (0.01) (0.01) (1.75) T 0.185** 0.145** 1.068 -0.068 -0.073 -65.299* (0.10) (0.08) (23.47) (0.02) (0.11) (24.75) T*HH schooling -0.028* -0.006 -1.328 -0.014** -0.007 -0.612 (0.01) (0.01) (1.90) (0.01) (0.01) (1.44) T 0.195* -0.053* -16.812* 0.165* -0.067** -16.408* (0.03) (0.03) (5.77) (0.03) (0.04) (5.70) T*HH is male -0.019 0.076** 35.752* -0.034 0.005 -1.361 (0.07) (0.04) (13.54) (0.07) (0.06) (10.54) T 0.165* -0.131* -51.842* 0.172* -0.082 -16.191 (0.06) (0.04) (12.70) (0.07) (0.06) (12.08) *Statistically significant at 5% level, **Statistically significant at 10% level (only for differences). Note: Expenditures levels are in Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/.13. Robust standard errors, clustered at the locality level. Sample includes 1359 households with observations in the panel 2000-2001-2002
33
Table 4: RPS Program Impacts along Observables Characteristics for children 7 to 13 years at baseline (continue) (Standard errors in parenthesis)
2001 2002 School
Attendance Participation
in labor activities
Hours Worked (Tobit)
School Attendance
Participation in labor activities
Hours Worked (Tobit)
Quintiles of Marginality Index T 0.190* -0.078** -24.558* 0.081* -0.072 -15.598 (0.08) (0.05) (11.17) (0.03) (0.08) (11.52) T*2nd Quintile -0.039 -0.001 2.590 0.112* -0.008 -3.149 (0.11) (0.06) (15.22) (0.05) (0.09) (15.89) T*3rd Quintile -0.122 0.100** 24.086 -0.018 0.075 9.348 (0.10) (0.06) (17.03) (0.05) (0.09) (15.85) T*4th Quintile -0.056 0.005 8.794 0.069 0.009 4.925 (0.09) (0.05) (12.04) (0.06) (0.11) (14.58) T*5th Quintile (richest) -0.012 0.017 4.535 0.107** -0.055 -13.667
` (0.11) (0.06) (13.35) (0.06) (0.09) (14.32) Quintiles of Household Per Capita Expenditures
T 0.197* -0.032 -13.496** 0.200* -0.062 -17.976* (0.06) (0.04) (8.38) (0.05) (0.04) (7.65) T*2nd Quintile -0.001 -0.058 -17.453 -0.030 -0.055 -4.258 (0.06) (0.05) (12.42) (0.06) (0.06) (10.06) T*3rd Quintile -0.033 0.014 6.149 -0.103 0.012 5.637 (0.06) (0.05) (11.30) (0.07) (0.06) (10.74) T*4th Quintile -0.147* -0.075** -14.530 -0.133** 0.010 7.082 (0.06) (0.04) (10.62) (0.07) (0.06) (10.29) T*5th Quintile (richest) -0.080 -0.040 -3.207 -0.031 -0.048 -2.381 (0.06) (0.06) (15.80) (0.07) (0.06) (10.90)
*Statistically significant at 5% level, **Statistically significant at 10% level (only for differences). Note: Expenditures levels are in Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/.13. Robust standard errors, clustered at the locality level. Sample includes 1359 households with observations in the panel 2000-2001-2002
34
Table 5: Tests of Rank Reversal from Distribution of Observables for Ranges in Expenditure Distribution, 2001
25q ≤ 25 50q ≤≺ 50 75q ≤≺ 75q Panel A: Per capita Total Expenditure Distribution Ranges
Mean Diff p-value Mean Diff p-value Mean Diff p-value Mean Diff p-value
Household Head is male 0.000 0.016 -0.001 0.609 -0.002 0.497 0.000 0.112 Household Head Years of Education 0.007 0.605 -0.007 0.345 0.004 0.050 0.011 0.058 Household Head is employed 0.000 0.856 0.002 0.441 -0.001 0.649 0.000 0.445 Household Head age -0.012 0.178 0.029 0.816 0.037 0.661 -0.003 0.455 Girls 0-5 years -0.003 0.567 0.001 0.527 0.001 0.972 -0.005 0.617 Girls 5-15 years 0.004 0.515 0.001 0.986 0.002 0.948 -0.009 0.816 Boys 0-5 years 0.005 0.948 0.000 0.513 -0.005 0.002 -0.006 0.178 Boys 5-15 years 0.003 0.820 0.001 0.341 -0.002 0.048 -0.009 0.074 Panel B: Per capita Food Distribution Ranges
Household Head is male 0.000 0.192 0.001 0.633 -0.004 0.503 0.000 0.150 Household Head Years of Education 0.008 0.583 -0.011 0.593 -0.009 0.140 0.011 0.130 Household Head is employed -0.001 0.463 0.004 0.439 -0.001 0.411 0.000 0.687 Household Head age -0.007 0.108 -0.083 0.471 0.171 0.749 -0.014 0.301 Girls 0-5 years -0.003 0.525 -0.001 0.549 0.003 0.926 -0.004 0.583 Girls 5-15 years 0.007 0.359 -0.002 0.864 -0.002 0.333 -0.009 0.443 Boys 0-5 years 0.004 0.317 0.004 0.523 -0.007 0.002 -0.005 0.062 Boys 5-15 years 0.003 0.497 -0.002 0.222 0.007 0.036 -0.010 0.098
35
Table 6: Tests of Rank Reversal from Distribution of Observables for Ranges in Expenditure Distribution, 2002 25q ≤ 25 50q ≤≺ 50 75q ≤≺ 75q Panel A: Per capita Total Expenditure Distribution Ranges
Mean Diff p-value Mean Diff p-value Mean Diff p-value Mean Diff p-value
Household Head is male 0.000 0.695 -0.001 0.371 0.000 0.629 -0.001 0.271 Household Head Years of Education 0.009 0.854 0.011 0.160 0.000 0.749 0.005 0.573 Household Head is employed 0.000 0.617 0.002 0.934 0.000 0.467 -0.001 0.495 Household Head age 0.107 0.391 -0.108 0.774 -0.036 0.383 0.061 0.291 Girls 0-5 years 0.006 0.982 -0.002 0.196 0.000 0.539 -0.007 0.379 Girls 5-15 years -0.003 0.022 0.011 0.112 0.004 0.389 -0.013 0.820 Boys 0-5 years 0.006 0.948 0.000 0.156 -0.001 0.056 -0.007 0.122 Boys 5-15 years 0.005 0.583 -0.002 0.028 0.003 0.944 -0.010 0.110 Panel B: Per capita Food Distribution Ranges
Household Head is male 0.000 0.365 0.001 0.695 -0.003 0.948 0.000 0.425 Household Head Years of Education 0.014 0.160 0.011 0.904 -0.010 0.902 0.006 0.471 Household Head is employed 0.000 0.665 0.002 0.788 -0.001 0.894 0.000 0.583 Household Head age 0.087 0.423 -0.075 0.170 -0.046 0.866 0.059 0.273 Girls 0-5 years 0.007 0.733 -0.002 0.443 -0.001 0.481 -0.006 0.435 Girls 5-15 years 0.000 0.032 0.012 0.255 0.000 0.084 -0.012 0.890 Boys 0-5 years 0.004 0.471 0.005 0.072 -0.004 0.361 -0.006 0.092 Boys 5-15 years 0.004 0.950 0.002 0.008 0.004 0.549 -0.011 0.248
36
Figure 1: Quantile Treatment Effect on the Distribution of
Annual Per Capita Total Expenditures in 2001
-100
400
900
1400
1900
2400
0 10 20 30 40 50 60 70 80 90 100
Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) In Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/13. (vi) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures.
37
Figure 2: Quantile Treatment Effect on the Distribution of
Annual Per Capita Total Expenditures in 2002
-100
400
900
1400
1900
2400
0 10 20 30 40 50 60 70 80 90 100
Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) In Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/13. (vi) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures.
38
Figure 3: Quantile Treatment Effect on the Distribution of
Annual Per Capita Food Expenditures in 2001
0
500
1000
1500
2000
2500
0 10 20 30 40 50 60 70 80 90 100
Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) In Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/13. (vi) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures.
39
Figure 4: Quantile Treatment Effect on the Distribution of
Annual Per Capita Food Expenditures in 2002
0
500
1000
1500
2000
2500
0 10 20 30 40 50 60 70 80 90 100
Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) In Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/13. (vi) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures.
40
Figure 5: Quantile Treatment Effect on the
Distribution of Food Share in 2001
-0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0 10 20 30 40 50 60 70 80 90 100
Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures.
41
Figure 6: Quantile Treatment Effect on the
Distribution of Food Share in 2002
-0.02
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0 10 20 30 40 50 60 70 80 90 100
Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures
42
Appendix Figure A.1: Quantile Treatment Effect on the Distribution of
Annual Per Capita Total Expenditures in 2000
-1200
-700
-200
300
800
1300
1800
0 10 20 30 40 50 60 70 80 90 100Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) In Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/13. (vi) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures.
43
Figure A.2: Quantile Treatment Effect on the Distribution of
Annual Per Capita Food Expenditures in 2000
-1200
-700
-200
300
800
1300
1800
0 10 20 30 40 50 60 70 80 90 100Quantile
Diff
eren
ces
Notes: (i) Solid line is the treatment quantile effect. (ii) Dashed lines provide confidence interval from the bootstrap with 1000 replications clustered at the locality level. (iii) Small dashed line is the mean impact. (iv) Sample includes 1359 households with observations in the panel 2000-2001-2002. (v) In Nicaraguan Cordobas, the equivalent exchange rate is $US1 = C/13. (vi) QTE is computed for the 91st to 99th quantiles but they are not included in the figures because their variances are large enough to distort the scale of the figures.
44
NOTES
* I am grateful for helpful comments and suggestions from Dan Black, Jeff Kubik, Jose
Galdo, Hugo Ñopo, the editor, and two anonymous referees. I am grateful to IFPRI for
permission to use the data. Any errors or omissions are responsibility of the author.
1 The RPS program did not identified poor households within targeted localities as in
PROGRESA. See Maluccio and Flores (2005) for an assessment of the targeting procedure.
2 Children are required to enroll and attend classes at least 85 percent of the time i.e. no more
than 5 absences every 2 months without valid excuse.
3 This design seems to embody a perverse incentive for students to keep repeating the 4th
grade so that families can continue to receive the subsidy. In order to eliminate this problem,
the program design included a number of causes for which the household may be expelled
from the program, among them, if the beneficiary child failed to be promoted to the next
grade. This condition, however, was deemed unfair and never enforced. Thanks to the referee
for pointing this out.
4 The lump sum transfer for school supplies and uniforms varies with the number of eligible
children while the school attendance transfer is a lump sum per household, regardless of the
number of children.
5 The evaluation was designed to last for one year, because of delays in funding the
implementation of the program was postponed in control localities until 2003.
6 In panel data, both non-response and attrition are potential concerns for the empirical
analysis. In 2001 and 2002, about 92 percent and 88 percent of the targeted households were
re-interviewed, respectively. The principal reasons for failure to interview targeted sample
households were that household members were temporarily absent or that the dwelling
appeared to be uninhabited. Maluccio and Flores (2005) examine the correlates of the
45
observed attrition and conclude that attrition is not a major concern for estimating program
effects and emphasize that using only the balanced panel is likely to slightly underestimate
the effects.
7 Maluccio and Flores (2005) analyze 15 indicators and find small differences only in
household size and number of children younger than 5 years old. Similarly, the
randomization is at the village level in Mexico’s PROGRESA, Behrman and Todd (1999)
found that treatment and control groups had similar mean outcomes at the locality level
before the program; however, they find small differences at the household and individual
level.
8 In the RPS data, approximately 20 percent of the beneficiary households had no targeted
children, 25 percent only children under age 5, 20 percent only children ages 7–13, and the
remaining 35 percent both under 5 year-olds and 7–13 year-olds.
9 The effect of RPS participation – the effect of “treatment on the treated” (ATT) – may also
be of interest. If the treatment effect on eligible non-participants is zero and if intent-to-treat
(ITT) is the overall impact effect evaluated at the sample mean, we can estimate the ATT by
rescaling the ITT by the fraction of program participants (Heckman, LaLonde, Smith, 1999).
In this experiment, however, the ITT is unlikely to underestimate the effect on participating
households by much as Maluccio, Murphy and Regalia (2006) show in their analysis of
impacts of the RPS program on schooling outcomes, because about 90% of eligible
households participate in the program.
10 Thanks to the referee for pointing this out. Unreported regressions show that the QTE
estimation without controlling for covariates are in the 95 percent range of the QTE estimates
controlling for covariates. With non-experimental data, the estimation can adjust for
differences in baseline observable characteristics by using propensity score weighting as in
Bitler et al. (2006) and Firpo (2007).
46
11 See Maluccio and Flores (2005) for more detailed information on the constructed variables.
12 Labor laws in Nicaragua establish age fourteen as the basic minimum age for work.
Children between the ages of 14 and 17 can work a maximum of six hours per day but not at
night. The employment of youth is prohibited in places that endanger their health and safety
such as mines, garbage dumps, and night entertainment venues (i.e. nightclubs, bars, etc.).
Government enforcement, however, is far from strict.
13 The bootstrap samples are drawn in a manner that mimics the stratified cluster sample
design of the RPS survey. The randomization was made at the comarca level, i.e. all
households in the treatment comarca are assigned to the program. The standard errors were
calculated by using a one-level bootstrap, that is, within each comarca J households are
randomly drawn. QTEs are calculated for each bootstrap sample and the process is repeated
1,000 times. The standard deviation of a QTE over the bootstrap replications is an estimator
of the standard error.
14 Figure A.1 and A.2 show the QTE estimates for the year before random assignment. The
structure of the figures is identical to before; they show the mean effect, the QTEs, and the
bootstrap 95% confidence interval of the QTEs. The effects are statistically not significant
different from zero for all quantiles.
15 See Bitler et al. (2005) for a detailed explanation of the joint test for the significance of the
demographic variables.